SoL-Pi vs token-optimizer
NVIDIA's Pi extension with four auto-research-discovered efficiency mechanisms - action fusion, observation packing, evidence-preserving log reduction, online context compaction - all opt-in. — versus — Context-waste hunter for coding agents: hooks compress reads, bash and search output, checkpoint before compaction, and audit waste in configs, skills, MCP and memory — with a local dashboard.
token-optimizer uses hooks to compress reads/bash/search output and checkpoint before compaction for Claude Code; SoL-Pi does the equivalent inside Pi.
| SoL-Pi | token-optimizer | |
|---|---|---|
| Stars | 2.3k | 2.3k |
| Forks | 181 | 176 |
| Language | TypeScript | Python |
| License | MIT | NOASSERTION |
| Last activity | yesterday | 2 days ago |
| Topics | coding, agents | coding, skills |
| Curated connections | 6 | 8 |
SoL-Pi — the curator's take
Install SoL-Pi if you run Pi as your coding agent and want fewer turns and less context replay without the agent skipping work: Action Fusion runs an edit's validation command in the same tool call, ObservationPack turns repeated big tool results into paged handles, the Reducer only compacts a log when every retained quote still matches the archive. It is an extension over unmodified Pi 0.85.1 - no patches, everything disabled until you write sol-pi.json. Not for Claude Code/Codex users (it uses Pi's extension APIs); headroom, rtk and token-optimizer are the harness-agnostic ways to get the same effect. The paper is the interesting part: the mechanisms were found by scaled auto-research loops over the harness itself, which is autoresearch/evo pointed at agent efficiency.
token-optimizer — the curator's take
The one token tool that argues with the others in its own README, and mostly wins the argument. Compressors like Headroom and RTK cover command output — roughly 15-25% of your context. Token Optimizer covers eight surfaces (bash, grep, tabular, file-re-read diffs, structure skeletons, archived large results, model verbosity, structural context) and then keeps going: checkpoints before auto-compact so savings survive it, model-routing nudges, loop detection, 30-day trend coaching, and per-component audits of CLAUDE.md, skills and MCP. It is cache-safe, injects nothing into your context, and measures before/after. The costs are real too: it is a large Python/TypeScript surface with 92 env knobs and 15 SQLite tables doing hook surgery on every Read and Bash, its headline dollar figures are counterfactual models against the author's own frozen baseline, and the license is non-standard. Pick RTK if you want one boring binary; pick this if you want the whole waste budget attacked and instrumented.